Multi-Method Diagnosis of Blood Microscopic Sample for Early Detection of Acute Lymphoblastic Leukemia Based on Deep Learning and Hybrid Techniques

Abstract

Leukemia is one of the most dangerous types of malignancies affecting the bone marrow or blood in all age groups, both in children and adults. The most dangerous and deadly type of leukemia is acute lymphoblastic leukemia (ALL). It is diagnosed by hematologists and experts in blood and bone marrow samples using a high-quality microscope with a magnifying lens. Manual diagnosis, however, is considered slow and is limited by the differing opinions of experts and other factors. Thus, this work aimed to develop diagnostic systems for two Acute Lymphoblastic Leukemia Image Databases (ALL_IDB1 and ALL_IDB2) for the early detection of leukemia. All images were optimized before being introduced to the systems by two overlapping filters: the average and Laplacian filters. This study consists of three proposed systems as follows: the first consists of the artificial neural network (ANN), feed forward neural network (FFNN), and support vector machine (SVM), all of which are based on hybrid features extracted using Local Binary Pattern (LBP), Gray Level Co-occurrence Matrix (GLCM) and Fuzzy Color Histogram (FCH) methods. Both ANN and FFNN reached an accuracy of 100%, while SVM reached an accuracy of 98.11%. The second proposed system consists of the convolutional neural network (CNN) models: AlexNet, GoogleNet, and ResNet-18, based on the transfer learning method, in which deep feature maps were extracted and classified with high accuracy. All the models obtained promising results for the early detection of leukemia in both datasets, with an accuracy of 100% for the AlexNet, GoogleNet, and ResNet-18 models. The third proposed system consists of hybrid CNN-SVM technologies, consisting of two blocks: CNN models for extracting feature maps and the SVM algorithm for classifying feature maps. All the hybrid systems achieved promising results, with AlexNet + SVM achieving 100% accuracy, Goog-LeNet + SVM achieving 98.1% accuracy, and ResNet-18 + SVM achieving 100% accuracy.

Keywords: acute lymphoblastic leukemia; convolutional neural network; fuzzy color histogram; gray level co-occurrence matrix; hybrid method; local binary pattern; machine learning.

Figure 1

Morphological diagram for leukemia diagnosis from the ALL_IDB1 and ALL_IDB2 dataset.

Figure 2

Samples from the (a) ALL-IDB1 and (b) ALL-IDB2 datasets.

Figure 3

Samples from the (a) ALL-IDB1 and (b) ALL-IDB2 datasets after enhancement.

Figure 4

Samples from the two datasets after segmentation.

Figure 5

Samples from the two datasets before and after the morphological processes.

Figure 6

Fusion of features extracted by the LBP, GLCM, and FCH algorithms.

Figure 7

Architecture of the artificial neural network and feed forward neural network classifiers of the ALL-IDB dataset.

Figure 8

AlexNet network structure for ALL-IDB dataset diagnostics.

Figure 9

GoogLeNet network structure for ALL-IDB dataset diagnostics.

Figure 10

ResNet-18 network structure for ALL-IDB dataset diagnostics.

Figure 11

Deep learning–machine learning hybrid techniques: (a) AlexNet + SVM; (b) GoogleNet + SVM; and (c) ResNet-18 + SVM.

Figure 12

Training of the feed forward neural network classifier of the ALL_IDB1 and ALL_IDB2 datasets.

Figure 13

Performance of the artificial neural network classifier of the (a) ALL_IDB1 and (b) ALL_IDB2 datasets.

Figure 14

Gradient and validation check values of the (a) ALL_IDB1 and (b) ALL_IDB2 datasets.

Figure 15

Receiver operating characteristic values of the (a) ALL_IDB1 and (b) ALL_IDB2 datasets.

Figure 16

Error histograms of the (a) ALL_IDB1 and (b) ALL_IDB2 datasets.

Figure 17

Regression of the (a) ALL_IDB1 and (b) ALL_IDB2 datasets.

Figure 18

Confusion matrix for the ANN algorithm of the (a) ALL_IDB1 and (b) ALL_IDB2 datasets.

Figure 19

Confusion matrix for the FFNN algorithm of the (a) ALL_IDB1 and (b) ALL_IDB2 datasets.

Figure 20

Performances of all the ANN, FFNN and SVM algorithms for the ALL_IDB1 and ALL_IDB2 datasets.

Figure 21

Performances of all the convolutional neural network models for the ALL_IDB1 and ALL_IDB2 datasets.

Figure 22

Confusion matrix for the AlexNet model for the (a) ALL_IDB1 and (b) ALL_IDB2 datasets.

Figure 23

Confusion matrix for the GoogLeNet model for the (a) ALL_IDB1 and (b) ALL_IDB2 datasets.

Figure 24

Confusion matrix for the ResNet-18 model for the (a) ALL_IDB1 and (b) ALL_IDB2 datasets.

Figure 25

Performances of all the hybrid techniques for the ALL_IDB1 and ALL_IDB2 datasets.

Figure 26

Confusion matrix for the AlexNet + SVM hybrid technique for the (a) ALL_IDB1 and (b) ALL_IDB2 datasets.

Figure 27

Confusion matrix for the GoogLeNet + SVM hybrid technique for the (a) ALL_IDB1 and (b) ALL_IDB2 datasets.

Figure 28

Confusion matrix for the ResNet-18 + SVM hybrid technique for the (a) ALL_IDB1 and (b) ALL_IDB2 datasets.

Figure 29

Comparison of the performances of the proposed systems for diagnosing leukemia from the two datasets.

Figure 30

Comparing the performance of our proposed system with the previous systems.